Schedule
Every 30 minutes
Model
OpenAI GPT-4o-mini
openrouter/openai/gpt-4o-mini
Approval
Every draft requires your approval before send
Outreach is queued in outbox/<domain>__<date>.md. You read, edit, click approve. Nothing leaves until then.
What it does
Every 30 minutes, scans for new prospect intake (a URL or company name). Runs a real recon via ollagraph: scrapes the homepage, checks WHOIS, pulls structured schema, extracts the language they actually use. Drafts a 90-word AIDA cold email referencing a real hook (a launch, a job posting, a public announcement) — never fabricated. Drops it in `outbox/` for your review. Nothing sends without you clicking approve.
Tools it can use
ollagraph
mcp_ollagraph_scrape Clean markdown of the prospect's homepage mcp_ollagraph_intel_whois Domain ownership + registration date (filters fake/throwaway domains) mcp_ollagraph_extract_structured Real legal entity name, location, schema.org data mcp_ollagraph_seo_keyword_extract The language they use about themselves — re-used in outreach filesystem
file_write Drop drafts in outbox/ memory
memory_store Track which prospects we've already targeted Workspace
Paths under /var/lib/openfang/workspace/ this Agent reads or writes:
outbox/shared.outbound.* Starter prompts
Recon a prospect
URL → 2-line summary
Recon: https://example.com — who they are, when their domain was registered, one outreach hook based on real recent activity. No outreach draft yet.
Draft cold outreach
After recon — full AIDA email
Draft a 90-word cold outreach email to the COO of <company>. Use the hook from the prior recon. AIDA structure, soft CTA, single-line signature.
Multi-touch sequence
Initial + 2 follow-ups
Plan a 3-touch outreach sequence (initial + follow-up 1 at +4d + follow-up 2 at +9d + breakup at +14d) for a prospect at Series B fintech. Each touch must add new value.
Disqualify a prospect
Stop wasting energy on no-fits
Quick disqualify check on https://example.com — is this prospect worth pursuing? Look at company stage, real product, current hiring signals. Recommend pursue / park / drop in 2 lines.